Title:

High-Performance Simulation of Bioelectric Fields

Field lines inside a human torso	Field density inside a human head
(Courtesy Chris Johnson, SCI, Utah)

Background:

The localization of bioelectric signals plays an important role in medical application, especially in Cardiology, Neurology and Neurosurgery. In the field of neurology the electroencephalogram (EEG) is the method of choice to get information on pathological electrical brain activity.

Compared to the classical EEG the model-based visualization of bioelectric fields has a much greater force of expression. One distinguishes two types of problems. In the so-called forward problem, the strengths and locations that are responsible for the bioelectric fields are known and one wants to compute the resulting potentials. In the related inverse problem the reconstruction of the sources from EEG measurements is the goal. For realistic simulations one has to take into account the individual patient anatomy. This leads to large systems of linear equations whose numerical solution is computationally very expensive.

Task:

This work focuses on efficient implementations of iterative solution methods for such large systems of linear equations. All modern workstations that are based on RISC processors have a hierarchical memory structure, which in general comprises two or three levels of fast cache memories. Data kept in cache memories can be accessed significantly faster than data stored in main memory or even on the hard disk. Only if the numerical codes are designed in a cache-aware manner a high efficiency in terms of MFLOPS can be expected. Usually, compilers are not smart enough to do a good job in optimizing these codes. Therefore, appropriate optimization techniques, like data access transformations and data layout transformations, have to be coded by the programmer.

The goal of this master's thesis is to experiment with techniques that can improve the implementation of iterative solution techniques, like e.g. SOR, multigrid methods and the method of conjugate gradients.

Contact:

Working on this project is Markus Zetlmeisl, his advisors are Marcus Mohr and Markus Kowarschik.